Allen Institute for AI released Olmo-core 3, a development framework intended to make mixture-of-experts large language model training more efficient. It reported 52,000 tokens per second on Nvidia B3000 GPUs for a 47-billion-parameter model, about 2.7× higher than Megatron-core at roughly 19,400 tokens per second. The update supports growing expert pools from 8 to 128 while selecting 4 experts per token, enabling scaling to over 1 trillion parameters with lower memory and compute overhead.
Amazon SageMaker AI fine-tuned a Qwen3.6-27B search agent using multi-turn reinforcement learning to improve how it searches across multiple interaction rounds. The failure rate on BrowseComp-Plus fell from 22.89% to 0.68%. Multi-turn RL raised retrieval quality on three of four held-out benchmarks and made the agent substantially more reliable by reducing task-completion errors within the turn and token limits.
Asta open-sourced AstaBrief, an 8B model that generates cited scientific reports from a user question and retrieved literature excerpts. The company reported Fast mode averaging 51.1 seconds per report versus 178.5 seconds for Thinking mode. As a result, AstaBrief is available to download and run on institutions’ own infrastructure, with released training data and an example workflow for local PDF-to-report generation.
Open-sourcing AstaBrief, a fast report-generation model built in Asta for producing cited scientific reports from research questions and retrieved excerpts.
ServiceNow CoreAI built AutoSynthData to turn enterprise agent capability gaps—identified from a target model’s environment failures—into executable training tasks generated and validated using a stronger teacher’s successes. AutoSynthData’s EnterpriseOps Gym pipeline uses a task-selection rule that favors target-model solves on no more than 1 of 3 trials while requiring the stronger solver to succeed on at least 2 of 3 trials. As the model improves, AutoSynthData shifts the curriculum toward tasks the model still struggles with, using iterative evaluation and feedback loops to expand and rebalance the training dataset.
A tutorial implements Google Research’s Kauldron JAX training library by demonstrating konfig (plain dict configs), kontext (string-key wiring that prevents components from importing each other), ktyping (runtime named-axis shape checks), and kd.train (a readable Trainer), then runs a synthetic CPU training run and a 5-variant config sweep with checkpointing and resume. It installs Kauldron pinned to version 1.4.2 and applies a small compatibility patch because a Trainer can otherwise raise AttributeError before completing a single step. The result is a workflow where changing one config line reshapes dependent values via references and where training components and tensor shapes are validated and rewired through data and strings instead of code edits.
Multilingual HuBERT-style self-supervised speech models improved linguistic learning when pretraining was made to better discriminate between languages. Phone-ABX error dropped from 11.6% in the bilingual baseline to 10.4% (monolingual: 10.8%) while lexical (sWUGGY) and prosodic performance also increased. The gains were largest when language discrimination was introduced in the first training iteration, whereas later or repeated use caused more language-wise segregation.
Every AI story that matters,
in your inbox by 8am.
TLDRocket reads all relevant sources, removes duplicate coverage, and summarises the
day in two minutes. Follow companies and topics for alerts, or get the
briefing in Slack. Free, no spam, unsubscribe anytime.
Reading TLDRocket needs no cookies, and the readership counts we rely on come from
our own cookieless analytics. Google Analytics is the exception: it sets cookies and
reports to Google, so it stays switched off until you allow it. You can change your
mind any time from “Cookie settings” in the footer.
Strictly necessary
Session security and form protection (tldrocket-session,
XSRF-TOKEN, 2 hours). The site cannot work without them,
so they need no consent.
Always on
Google Analytics 4 (_ga,
_ga_<id>, up to 2 years). Measures which
stories and sections readers use. Google acts as a third-party processor and may
store the data outside the EU. No advertising, no profiling, no data sold.
